2026-03-08

Measuring the Pipeline

V2.0 made the pipeline visible. v2.1 asks whether it’s correct. We rebuilt five classic human vision experiments as automated reference pages, ran the shader against each one, and compared pixel measurements to the published data. The whole battery (stimulus pages, capture scripts, analysis, golden captures) ran in a single day. Three bugs (two in the shader, one in the analysis) fell out that prior testing had not caught.

Chromatic decay: Mullen 2002, Hansen 2009, Bowers 2025
Spatial frequency: Rovamo & Virsu 1979
Crowding geometry: Bouma 1970, Toet & Levi 1992
Saliency protection: Itti & Koch 2001, Hershler 2005
Mixed-density search: Halverson & Hornof 2011

Also in this release: 8 half-octave DoG bands replace the old 4-band decomposition. Smoother blur gradient, twice as many frequency transitions, same API.

Five waves of validation

Each wave targets a different pipeline stage. The pattern is the same throughout: render a known stimulus through the shader, measure output pixels, compare against published human data.

Four experimental stimuli used for psychophysical validation: color search singletons, spatial frequency gratings, flanked crowding letters, and saliency pop-out targets — **The four stimulus classes.** Each recreates a published psychophysical experiment as an HTML reference page. Top-left: crowding geometry (Bouma 1970). Top-right: spatial frequency at 1 cpd (Rovamo & Virsu 1979). Bottom-left: chromatic decay (Hansen 2009). Bottom-right: saliency with face detection (Itti & Koch 2001; Hershler & Hochstein 2005).

The same four stimuli rendered through Scrutinizer's peripheral vision pipeline, showing eccentricity-dependent degradation — **Pipeline output.** The stimuli rendered through the full shader at center fixation. High-frequency content (fine gratings, letter serifs) degrades first. Chromatic content undergoes eccentricity-dependent pooling. Bottom-right uses the pop-out with modulation variant (color singleton + luminance squares). Pixel measurements at each eccentricity ring are compared against the published human data.

Wave	Domain	Published basis	Finding
1	Chromatic decay	Mullen 2002, Hansen 2009, Bowers 2025	Green tracks the RG curve, not BY; hue-based models get this wrong
2	Spatial frequency	Rovamo & Virsu 1979	DoG produces step functions at MIP boundaries vs smooth CSF
3	Crowding geometry	Bouma 1970, Toet & Levi 1992	R:T bug found and fixed; density gate validated at 3.3:1 ratio
4	Saliency protection	Itti & Koch 2001, Hershler 2005	Face saliency 4.79× control; protection ratio 0.283
5	Mixed-density search	Halverson & Hornof 2011	Density gate validates at macro level; word-level granularity gap found

The full validation infrastructure ships with the repository: 5 capture scripts, 5 analysis scripts, 15 reference pages, 25+ golden captures, and machine-readable JSON of published data spanning 46 years of vision science (Rovamo 1979, Mullen 2002, Hansen 2009, Bowers 2025). Every claim in the arxiv paper can be reproduced by running the scripts.

What the numbers say

Each wave is graded at three tiers: must-pass (correct qualitative behavior), should-pass (quantitative agreement), stretch (publication-quality correlation with human data).

Wave 1 (chromatic decay): 5/5 Tier 1, 3/3 Tier 3. Blue retention at ring 5 is 1.70× red, matching the Mullen & Kingdom (2002) RG/BY asymmetry. Model retention correlates $r = 1.000$ with Hansen et al.’s (2009) color naming data. Green retains 51.9%, closer to red (43.1%) than blue (73.5%). In Oklab, green’s chroma splits ~57% RG / ~43% BY. As the RG channel collapses peripherally, green shifts toward yellow-green. Opponent-channel models predict this shift from the channel split. Hue-based models do not.

Update (2026-10-02): The Hansen et al. (2009) color naming data cited in this post did not come from that paper, which reports chromatic detection thresholds at 10–50°. The data file and its Tier 3 check have been removed. The check was a rank correlation of two curves that both fall with eccentricity, so it returned r = 1.000 for any monotone model and could not fail. Wave 1 Tier 3 now has one check: blue-yellow retention above red-green at every ring.

Wave 2 (spatial frequency): 11/11 Tier 1, 5/5 Tier 2, 0/1 Tier 3. Frequency ordering is correct everywhere, M-scaling cutoffs match exactly, and the residual band (layout blocks) survives at 100% across all rings. But Tier 3 fails: per-band Spearman correlations with Rovamo & Virsu (1979) range from $r = -1.0$ to $r = 0.6$ because each DoG band transitions 100%→0% in a single step. The composite metric is $r = 0.600$. The 8-band upgrade narrows each step but doesn’t eliminate the underlying discretization. A matched Gaussian blur comparison confirmed the root cause: DoG and Gaussian produce near-identical contrast retention slopes on achromatic gratings (ring 4: both $-0.4675$; ring 5: $-0.3592$ vs $-0.3601$). Both pipelines sample from the same MIP chain, so the band decomposition doesn’t add frequency selectivity beyond what the MIP pyramid already provides. The differentiation is architectural: saliency gating, chromatic channel separation, and density-gated crowding are pipeline stages that Gaussian blur lacks entirely, and the saliency comparison shows 5–10× better content preservation.

Update (2026-10-01): The 5–10× figure overstates the saliency effect. The control region, which has no saliency signal, also shows a 3.2× advantage, so the Gaussian condition was not matched in total information loss. Only the excess over the control, about 1.8–3.4×, can be attributed to saliency gating.

Three bugs found by measurement

Polar sector R:T ratio: spoke count computed from biased ring width, producing ~1:1 instead of the 2:1 Toet & Levi measured. Fixed.
V1 far-peripheral growth: displacement plateaued beyond the parafovea. Growth factor tuned from 0.5 to 1.5.
Rovamo composite metric: per-band correlations were meaningless for step functions. Replaced with frequency-weighted composite.

The Halverson gap

Halverson & Hornof (2011) found peripheral encoding accuracy drops from 90% to 50% when word spacing falls below 0.15°. Scrutinizer’s density gate makes the same structural prediction (distortion strength scales with density via a sigmoid), but the structure map works at the level of DOM blocks. It does not resolve individual words. A sparse group (5 words, 0.65° spacing) and a dense group (10 words, 0.33°) produce nearly identical density values. SSIM difference: 0.3% at 90px foveal radius, 0.1% at 60px. The gate validates at the macro level (sidebar vs hero image). It fails at the micro level (word spacing within a block). Congestion scoring (Rosenholtz Feature Congestion) provides a complementary signal computed from pixel-level texture statistics, which partially bridges this granularity gap.

Update (2026-10-01): The same test found that congestion does not bridge the gap: Mode 9 (congestion-gated) changed sparse/dense discrimination by only 0.1% relative to Mode 0. At a 38px (1°) foveal radius the sparse/dense difference ran the wrong way.

Known limits

Stepped blur. 8 bands are smoother than 4 but still change in discrete steps, where human vision has continuous falloff.
Crowding strength only. The density gate scales distortion intensity. It does not change the critical spacing at which crowding engages. The shader has no information about what is adjacent to a given pixel.
No stimulus-specific crowding. Same-orientation flankers should crowd harder than orthogonal ones (Pelli & Tillman 2008). The renderer has no concept of target-flanker similarity.

Update: v2.2’s oriented DoG bands partially address the third limitation. Cardinal-aligned edges (horizontal, vertical) now receive longer M-scaling cutoffs than oblique edges, and radial edges fade faster than tangential ones. This isn’t full stimulus-specific crowding (flanker similarity is still not computed), but it means orientation now modulates degradation strength.

8 half-octave DoG bands

The previous version decomposed the image into 4 frequency bands: fine detail, medium detail, coarse structure, layout. The blur transition from sharp to pooled had 4 steps, which meant the parafoveal region (where text goes from readable to unreadable) had only 1–2 active bands at any point.

v2.1 doubles the resolution to 8 bands at half-octave spacing. The blur gradient is now twice as smooth, with the biggest improvement in the critical parafoveal zone. The old 4 bands are still there (bands 1, 3, 5, 7 below), and the new bands interleave between them.

Band	Freq (cpd)	Content	New?
0	5.66	Serifs, fine detail	New
1	4.0	Thin strokes	= old band 0
2	2.83	Letter bodies	New
3	2.0	Small icons	= old band 1
4	1.41	Words, UI labels	New
5	1.0	Word groups	= old band 2
6	0.71	Buttons, panels	New
7	0.5	Layout blocks	= old band 3

The performance cost is under 1ms at 1080p. See the MIP chain explainer for how the band decomposition maps to hardware texture levels.

Also in this release

See the full release notes for details on all changes, including 15 open-source experimental stimuli, arxiv paper updates (Walton 2021, WebGPU tiered roadmap, mongrel Tier 2.5 spec), TEST_LOAD_TIMEOUT for heavy pages, and appendix baseline captures.

References cited: Bouma 1970 · Bowers, Gegenfurtner & Goettker 2025 · Halverson & Hornof 2011 · Hansen et al. 2009 · Hershler & Hochstein 2005 · Itti & Koch 2001 · Mullen & Kingdom 2002 · Rovamo & Virsu 1979 · Pelli & Tillman 2008 · Toet & Levi 1992 · Annotated Bibliography · Primer References